IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO.

12, DECEMBER 2005

1529

Intrathoracic Airway Trees: Segmentation and Airway

Morphology Analysis From Low-Dose CT Scans
Juerg Tschirren*, Member, IEEE, Eric A. Hoffman, Geoffrey McLennan, and Milan Sonka, Fellow, IEEE

AbstractThe segmentation of the human airway tree from volumetric computed tomography (CT) images builds an important
step for many clinical applications and for physiological studies.
Previously proposed algorithms suffer from one or several problems: leaking into the surrounding lung parenchyma, the need for
the user to manually adjust parameters, excessive runtime. Lowdose CT scans are increasingly utilized in lung screening studies,
but segmenting them with traditional airway segmentation algorithms often yields less than satisfying results.
In this paper, a new airway segmentation method based on fuzzy
connectivity is presented. Small adaptive regions of interest are
used that follow the airway branches as they are segmented. This
has several advantages. It makes it possible to detect leaks early
and avoid them, the segmentation algorithm can automatically
adapt to changing image parameters, and the computing time is
kept within moderate values. The new method is robust in the
sense that it works on various types of scans (low-dose and regular
dose, normal subjects and diseased subjects) without the need for
the user to manually adjust any parameters. Comparison with a
commonly used region-grow segmentation algorithm shows that
the newly proposed method retrieves a signicantly higher count
of airway branches.
A method that conducts accurate cross-sectional airway measurements on airways is presented as an additional processing
step. Measurements are conducted in the original gray-level
volume. Validation on a phantom shows that subvoxel accuracy is
achieved for all airway sizes and airway orientations.
Index TermsAdaptive region of interest, airway tree segmentation, fuzzy connectivity, pulmonary imaging, quantitative analysis.

I. INTRODUCTION

HE knowledge about the location of the pulmonary airway

tree as well as its geometrical properties build an important step for the diagnosis, study, and treatment of lung disorders as well as for physiological studies of pulmonary function.
Low-dose computed tomography (CT) scans are increasingly
utilized in lung screening studies. Traditional region growing
segmentation frequently fails due to a high noise content of
low-dose images.

Manuscript received April 25, 2005; revised August 15, 2005. This work
was supported in part by the National Institutes of Health (NIH) under Grant
HL-064368. The Associate Editor responsible for coordinating the review of
this paper and recommending its publication was B. van Ginneken. Asterisk indicates corresponding author.
*J. Tschirren was with the Department of Electrical and Computer Engineering, The University of Iowa, Iowa City, IA 52240 USA. He is now with
VIDA Diagnostics, 100 Oakdale Campus, 19 TIC, Iowa City, IA 52242 USA
(e-mail: juerg@vidadiagnostics.com).
E. A. Hoffman and G. McLennan are with the Department of Radiology, The
University of Iowa, Iowa City, IA 52240 USA.
M. Sonka is with the Department of Electrical and Computer Engineering, The University of Iowa, Iowa City, IA 52240 USA (e-mail:
milan-sonka@uiowa.edu).
Digital Object Identier 10.1109/TMI.2005.857654

The work presented in this paper is divided in two parts,

airway segmentation and quantitative measurements. The purpose of the airway segmentation step is to determine the location of segments of the intrathoracic airway tree from CT image
volumes. The airway tree segmentation does not return precise
geometrical informationan accurate delineation of the airway
wall is performed by the quantitative analysis step.
Previous work on airway segmentation mainly includes
region growing-based methods [1][4], morphology-based
methods [5][7], and combinations of the two [8][10]. Other
methods proposed in the past include rule-based methods [11],
[12], energy function minimization [7], and region of interest
(ROI) modication-based techniques [13]. Schlathlter et al.
[14] use a front-propagation algorithm for segmenting airway
trees. Branchpoints are detected when the front splits up.
Diameters are measured during the segmentation and a leak is
identied if the diameter suddenly increases.
There are many reasons why airway tree segmentation is
difcult to achieve from volumetric CT images in a robust
fashion from clinical-quality or low-dose CT images. Some
of the reasons are anatomy-related, e.g., airway obstructions;
others are caused by heart beat induced movement artifacts
and image reconstruction artifacts. Consequently, one of the
biggest problems when segmenting airway trees with automated methods is leakage into the extra-luminal regions. Leaks
occur because the airway-wall that separates the airway lumen
from the surrounding lung parenchyma is relatively thin. The
lung parenchyma, on the other hand, exhibits a texture similar
to that of small airways. Partial volume effect and image noise
greatly decrease the CT visibility of the airway-wall. At some
places this can cause the segmentation algorithm to grow
through the airway-wall. Once this happens, potentially large
regions of lung-parenchyma are wrongly identied as airways.
Another common problem with previously proposed airway
segmentation algorithms is their difculties with segmenting
low-dose scans and scans of heavily diseased lungs, for example, lungs of emphysema patients. In the case of low-dose
scans, the segmentation either stops early or leaks. The user then
has to run the segmentation algorithm several times in the attempt to nd an optimal combination of parameters. In the case
of diseased lungs, heavy leaking is not unusual.
The quantitative measurement step follows the segmentation
step, and its purpose is to accurately nd the location of the inner
airway wall in order to conduct geometrical measurements. The
medical community expresses a great interest in not only the
qualitative but also the quantitative in particular also the quantitative analysis of airways [15].

0278-0062/$20.00 2005 IEEE

1530

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO. 12, DECEMBER 2005

In the past, measurements have often been done manually

[16], [17]. Automated methods [6], [18][22] have been presented before. Reinhardt et al. [18], [19] showed that the accuracy of measurements depends on the size of the airway and
proposed a model-based system using a twodimensional (2-D)
Gaussian function to model the point-spread function of the
scanner. Prteux et al. [20], [6] approximated the inner airway
wall with a Butterworth polynomial, but concluded that their
method produces a sharply increased error for small airways.
Saba et al. [21] used the full-width at half-maximum method to
conduct measurements in 2-D and then model the three-dimensional (3-D) airways based on these measurements.
Our new airway segmentation method reported below offers
fully automated and reliable identication of airway trees from
CT volumes. The need for human operators to manually optimize segmentation parameters was eliminated. The algorithm
works without a change on different types of scans, e.g., lowdose and regular-dose, on diseased subjects and normal subjects
while the runtime does not exceed several minutes per volume.
The quantitative airway size measurements step provides accurate airway dimensions (minor and major diameter, cross-sectional area) on airway segments of all sizes and all possible spatial orientations (including in-plane, relative to the scanning direction). Again, the measurements are performed fully automatically, with no manual intervention.

II. METHODSAIRWAY TREE SEGMENTATION

A. Overview
The segmentation algorithm presented here is based on fuzzy
connectivity as proposed by Udupa et al. [23] and Herman et al.
[24].
During the execution of this algorithm, two regionsforeground and backgroundare competing against each other.
This method has the great advantage that it can overcome
image gradients and noise. The disadvantage is its relatively
high computational complexity. Computing time can be reduced by splitting the segmentation space into a number of
smaller subspaces and by keeping the search space as tight as
possible around the airway segments.
As mentioned above, leaks into the surrounding lung
parenchyma are a common problem of previously proposed
airway tree segmentation algorithms. The root of a leak is a
very localized phenomenon. The leak itself, the location where
the segmentation result enters the lung parenchyma, normally
occurs within a very small area (within a few voxels). But if the
leak goes undetected then it often spreads quickly, eventually
occupying substantial parts of the lung and, in many cases,
rendering the airway tree segmentation useless (see Fig. 1).
The desire to keep the segmentation within a small area, together with the need to detect possible leaks at their root, lead
to the idea of using a relatively small adaptive ROI that follows
the airway tree branches as they are segmented. The ROI has a
cylindrical shape and adapts its geometrical dimensions, its orientation, and position to the predicted size, orientation, and position of the airway branch to be segmented. This has two main
advantages:

Fig. 1. Example of a severe segmentation leak. (Emphysema patient,

segmented with standard 3-D region-grow algorithmleak was unavoidable).

The segmentation process is kept close to the airway segments; therefore, the individual problem size is kept small,
which leads to faster segmentation time.
Problems (leaks) can be detected early and addressed.
Fig. 2 illustrates the concept. Using a cylindrically shaped ROI
(versus the more common cubical ROI used in other 3-D image
segmentation tasks) has the advantage that the ROI better adapts
to the target shape, which is close to cylindrical. This means less
useless background voxels have to be analyzed and the computing time can be shortened. A similar approach was independently used by Mori et al. [13].
B. Multiseeded Fuzzy Connectivity
The basic idea of segmentation with fuzzy connectivity is that
the voxels of an input image are compared with a seed-voxel
and the similarity/dissimilarity is expressed as a fuzzy membership value. The similarity of two voxels and is expressed by
, which is normally computed
the afnity value
based on gray-values and dened for adjacent voxels only. In
this application, we consider 18-connected voxels as adjacent.
If and are not directly adjacent then their similarity is compared by looking at all possible chains of adjacent voxels that
connect and . The strongest chain is chosen to represent the
similarity. The strength of a chain is dened by the lowest
value along its length (weakest link). Voxels are assigned to the
foreground region if their value exceeds a predened value.
The multi-seeded fuzzy connectivity (MFC) method takes
this idea one step further by growing two regions (foreground
and background) simultaneously and letting them compete for
voxels. The method guarantees that both resulting regions, foreground and background, are connected in themselves, i.e., no
isolated islands may occur. The advantage of segmentation
with MFC is that it can overcome image gradients and image
noise without signicant leaking or premature abortion of segmentation. The disadvantage of the method is that it is relatively expensive to compute, particularly for big 3-D volumes.
Methods for increasing computational efciency are presented
below.

TSCHIRREN et al.: SEGMENTATION AND AIRWAY MORPHOLOGY ANALYSIS FROM LOW-DOSE CT SCANS

1531

Fig. 2. Basic concept of airway tree segmentation. Adaptive cylindrical regions of interest (light gray) follow airway tree branches as the segmentation proceeds.
Segmentation is performed in a small area only, which keeps the computing time down. Possible problems (leaks) can be detected early and addressed. The
simplied ow diagram to the right does not show all details (for example, the termination criteria).

C. Input Data
To decrease the computer memory size requirements, the
obtained from the CT scanner
12-bit gray-scale values
using the piecewise linear
are converted into 8 bit values
function

(1)

Note that no information is lost in the range between

Hounseld Units (HU) and
HU, which includes all important information for the airway segmentation. This compression eases memory requirements and decreases computing time
due to the decreased memory requirements.
D. Afnity Functions
The airway lumen shows a mostly constant gray-level value
of 0 (measured in the 8-bit volume; corresponding to
HU) throughout the whole airway tree. Exceptions are noise
voxels, frequently seen in low-dose scans. An afnity function
based on the absolute value of the gray-value can be used. The
of a foreground voxel is solely based on the grayafnity
value of and is dened by
(2)
and
In the presented methods, values of
were determined experimentally and were kept constant across
all analyses. The exact values turned out to be not very criticalthe segmentation result is nearly identical over a relatively
and .
wide range of
The average gray-value of the airway wall changes substantially between different airway generations. For the rst generation (trachea), the gray-values of the wall voxels are mostly in
HU. But further down the tree, at the
the range of 0 to
5th or 6th generation, for example, the gray-values are in the
to
HU. This shift is caused by the partial
range of
volume effect introduced by the X-ray beam collimation and the
sampling process in the CT scanner to which the thinner airway

Fig. 3. Directional afnity. (a) Principle illustrated in 2-D. Gray-values of n

and n are included in computation of afnity between c and n . D marks
preferred direction. (b) Possible values for D in 3-D.

walls of the higher-generation airways are more sensitive. Motion artifacts and/or anatomical obstructions further complicate
the problem.
Because of the gray-value gradient, the afnity function
for the airway wall is based on the gray-value difference between two voxels and and is dened by
(3)
with
and

and

being the gray-values of

and

, respectively,

(4)
Note that the background corresponds to the airway wall and
that the parenchymal parts of the lung volume are not utilized
in the MFC segmentation.
E. Directional Afnity
With the afnity functions described above, some leaks are
unavoidable due to voxel-connections at the
HU level
between airway lumen and lung parenchyma. Such leaks are
mostly a very localized phenomena; the leaking bridge is normally only one or two voxels wide. Adding directional information to the afnity function helps prevent them.

1532

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO. 12, DECEMBER 2005

Fig. 4. Example of result from directional afnity. (a) New branch taking off to the left. (b) A leak occurred and is detected. (c) Previous step was deleted and by
using directional afnity a new leak can be avoided. (d) Segmentation of the branch continues with un-directed afnity function.

Directional afnity uses information about the expected spatial direction of the current airway segment. Fig. 3(a) illustrates
the concept in 2-D. When computing the afnity value between
a voxel and its neighbors , only neighboring voxels that lie
close to the axis of the expected direction D are considered, i.e.,
, and
in Fig. 3(b). All other neighbors get
voxels
an afnity value of zero assigned, relative to . Additionally,
the afnity value also depends on the gray values of further
neighbors in direction D. The directional foreground afnity
is computed with
(5)
The direction D is given by the spatial orientation of the ROI.
26 discrete directions are possible as illustrated in Fig. 3(b).
With a given ROI the closest discrete direction is used.
Fig. 4 shows an example where a leak was successfully
avoided with the help of directional afnity.
F. Positioning New Cylindrical Regions of Interest
The surface of the segmentation result in Fig. 5(a) is found
by a queue-based region-grow algorithm that only grows along
foreground-voxels that are in the 6-connected neighborhood
of at least one background-voxel. Additionally, the algorithm
distinguishes between surface voxels that are situated below
the surface of the ROI-cylinder, and voxels that lie on the
surface of the ROI. We call an isolated (8-connected) group of
the latter a surface face. Surface faces are used for initializing
the skeletonization process. The orientation of new ROIs is
determined based on the skeleton of the current segmentation
result. A rough skeletonization sufces for that purpose; exact
branch-point positions and smoothed skeletal lines are not
important. A computationally efcient skeletonization is implemented by rst computing the distance map of the segmentation
result. A simple graph search is then executed using the centers
of gravity of the surface faces as anchor points. If there are
more than two surface faces found along the surface of the ROI
then at least one branch-point must have been encountered.
The exact location of an ROI is not critical. The length of an
ROI may for example be expanded such that it partially covers
a neighbor airway branch. In that case, the neighbor branch will
still be explored independently since it was already detected
during the evaluation of the parent-ROI.

Fig. 5. Analysis of segmentation result. (a) Surface region growing result, with
surface faces colored light gray. (b) Skeletonized segmentation result.

An ROI may end in the middle of a branch-point. This case

is detected by an unusual increase in the area of the face at the

TSCHIRREN et al.: SEGMENTATION AND AIRWAY MORPHOLOGY ANALYSIS FROM LOW-DOSE CT SCANS

1533

downstream end of the ROI. In that case, the length of the

ROI is increased and the segmentation is repeated. This makes
sure the roots of the child-branches are detected correctly. The
default length of an ROI is 25 voxels and it can be lengthened
at most twice (by a factor of 2 each time).
G. Seedpoints
Two seedpoints are needed for the segmentation of each ROI:
one within the airway lumen and one within the airway wall.
For the initial ROI an additional seedpoint is required to determine the spatial orientation of the ROI. These initial three
seedpoints are automatically found by searching for the trachea.
In the transversal view, the trachea is a roughly circular object
that is located approximately in the center in the eld of view.
Further more the cross-sectional area is within a known range.
Taking this preknowledge the CT volume is thresholded slice
by slice and a circular object is sought that ts the description
above. If such an object can be found continuously over several consecutive slices then the trachea is found. This identies
the lumen- and the orientation-seedpoints. The wall-seedpoint
is identied by seeking the immediate periphery of the lumen
for a bright voxel.
The two seedpoints for every consecutive ROI are found
similarly. The lumen seedpoint is identied as the rst centerline voxel that lies within the new ROI, and the wall-seedpoint
is again found by scanning the lumens periphery for a bright
voxel.
H. Leak Detection
By closer examination of typical leaks, it was observed that
a leaked segmentation result almost always exhibits a spongy
structureit contains many holes and tunnels. A leak-free segmentation result, on the other hand, almost never shows these
properties. This observation was used to build a leak detector.
A sequential- and topology-preserving thinning [25] is applied
to the segmentation result. The size of the resulting topological
kernel is used as a leak detector. The topological kernel is the
object that remains after the sequential thinning step removed
all voxels that can be deleted without altering the topology of
the object. A topological kernel of size 1 (one single voxel) is
an indicator for a solid segmentation result
.A
indicates that a leak occurred. In
topological kernel of
other words, if for example the segmentation result contains a
hole then the thinning result (the topological kernel) will have
the shape of a ring. Naturally this ring must consist of more than
one voxel and this is taken as an indication of a leak.
I. Framework
Within the algorithm, the term frame refers to the data structure that contains all the information needed for segmenting one
ROI position. This includes the information about the position,
orientation, and size of the ROI itself, the parameters for the
foreground and background afnity functions, the position of
the seed points, the topological generation number of the current airway segment, and information about the current status
and the number of previously failed segmentation attempts at
this position.

Fig. 6. Quantitative analysis: 2-D slices are re-sampled perpendicular

to centerline, inner border is detected, cross-sectional area, minor-, and
major-diameter are computed.

Frames that are to be processed are put into a priority queue,

which uses the generation number of the airway segment as the
sort key. The dequeue function of the priority queue always returns the frame with the lowest generation number available in
the queue. The result is that the tree grows evenly. The segmentation of lower generation segments is nished before the
segmentation of any higher generation segments is attempted.
An advantage of this concept is that the number of segmented
airway generations can be controlled, which saves computing
time if only a few generations are to be retrieved.
If a segmentation problem (e.g., leak) is detected then the segmentation result of the last frame is deleted, the parameters of
the affected frame are modied, and the frame is re-queued. Possibly modied parameters include the geometry of the ROI (diameter, length, and position) and the afnity function (values
, and directionality). The kind of parameter changes defor
pends on the type of segmentation problem and is controlled
by a predened nite state machine. This nite state machine
has a depth of 3, i.e., if the segmentation of a particular frame
still fails after 3 parameter changes then this frame is given up
permanently and segmentation continues at a different airway
branch.
III. METHODSQUANTITATIVE ANALYSIS
The process of quantitative analysis of airway segments uses
the results of the previously-described segmentation of the
airway tree (Section II). Consequently, approximate surfaces of
the airway tree segments together with the airway tree skeleton
centerlines are used to guide the accurate detection of the
airway walls. The process is, thus, divided into the following
three steps (Fig. 6).
Re-sample 2-D slices perpendicular to airway segments.
For each 2-D slice separately, segment airway wall.
Conduct measurements on segmentation result.
is re-sampled from the original gray-level
A 2-D slice
volume using the centerline information, and a 2-D slice
is re-sampled from the earlier airway segmentation result. The

1534

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO. 12, DECEMBER 2005

is positioned in the center of the image). Simultaneously, the

same horizontal shift is applied to the corresponding line in the
radially re-sampled gray-image [Fig. 7(c), (d)]. The amount and
direction of shift as well as the radial position of every re-sampled point are recorded so that at the end of the segmentation
process the resulting border points can be mapped back into
used for the dythe original 2-D slices. The cost-function
namic programming uses the rst and second derivatives of the
gray-level image and is based on a cost function proposed in
[27]. Strictly speaking the cost-function used here should be
called the reward-function since higher values are preferred
over lower ones. But to be consistent with the literature the more
common term cost-function is used hereand the dynamic programming is maximizing the cost
(6)

Fig. 7. Radially re-sampled 2-D cross sections. (a) Preliminary

airway-segmentation
result
before
shifting.
(b)
Preliminary
airway-segmentation result after shifting. (c) Gray-level image before
shifting. (d) Gray-level image after shifting. The black frames around (a) and
(b) have been added for visualization purposes only; they are not part of the
original image.

slices are oriented perpendicular to the centerline of the respective airway segment and one pair of slices is re-sampled for
every centerline voxel position. Re-sampling at every centerline voxel position assures that every possible position along
the airway tree segments is covered by measurements. The perpendicular orientation is determined by computing the tangent
to the smoothed centerline (the centerline is obtained from the
skeletonization process). Re-sampling is performed using trilinear interpolation.
For every voxel along the centerline, the wall-segand
as input and outputs
mentation takes
with
as output, where
represent the segmented points along the inner airway
border, dened in the Cartesian coordinate system of
Dynamic programming [26] is used for the wall-segmentation.
The input images are radially re-sampled in order to stretch
the target border. Starting from the point dened by the centroid
, a total of
evenly
of the segmentation result in
points are
spaced rays are cast, and along each ray
sampled from
and
. This results in the images
and
of dimension
shown in Fig. 7(c) and (a).
The cross sections of airways are normally not perfectly circular, and as a result the airway-wall borders in Fig. 7(a) and (c)
are not completely straight after resampling. The straightness
of the border is, however, preferred for the graph search-based
segmentation so that the preliminary airway segmentation can
be used for guidance. Therefore, a local horizontal shift is applied to the radially re-sampled segmentation result shown in
Fig. 7(a) so that the straightened image resultsFig. 7(b) (the
transform is such that the transition between black and white

where symbolizes a 2-D convolution, the symbol stands for

is
smoothed with a 5 5
point-wise summation,
is a 5 5 Sobel mask, and
is a 5 5
unit matrix.
Marr mask. By changing the value of constant , the position
of the resulting border can be pushed or pulled radially as illustrated in Fig. 8. This is based on a well-known behavior of
rst- and second-derivative edge detectors that consistently position their maximum edge responses on one or the other side
of the true edge, respectively. Consequently, weighting their responses can be used for accurate positioning of the edge detection response with respect to the correct edge location. The
purpose of this is to adapt the cost-function to the estimated size
of the current airway segment and, thus, maintain a high degree
of measurement accuracy across all sizes of airways. The size
of the airway is estimated based on a pixel-count in the re-sam. The value of is determined via
pled segmentation result
the empirically determined piece-wise linear function shown in
Fig. 9. The values of were found by an optimization process
during which phantom CT images of tubes with known sizes
were used, the sizes covered the full range of airway sizes and
corresponded to the CT imaging parameters used for the in vivo
scans.
Measurements are conducted on the segmented border ,
where represents a polygon. The cross-sectional area is computed and the minor (minimal) and major (maximal) diameters
are found. The minor and major diameter lines are guaranteed to
pass through the center of gravity of the area inscribed by . The
results are written into an XML le that holds all tree-related
data (topology, geometry, anatomical labels, measurements) and
allows easy access to data for further processing.
IV. VALIDATION OF AIRWAY SEGMENTATION
A. Method
Validation of the airway segmentation would ideally be based
on a gold standard provided by a human expert who would perform hand-tracings of all discernible airway segments for several airway trees. Unfortunately this approach is not feasible due
to the extreme labor-intensity of the hand-tracing task.

TSCHIRREN et al.: SEGMENTATION AND AIRWAY MORPHOLOGY ANALYSIS FROM LOW-DOSE CT SCANS

1535

Fig. 8. Border position can be modied by adjusting the value of ! in (6). Note that the airway size is overestimated in (a) and underestimated in (c). (a) ! = 0:00,
(b) ! = 0:25, and (c) ! = 1:00.
TABLE I
COMPARISON OF NUMBER OF SUCCESSFULLY SEGMENTED NAMED AIRWAY
BRANCHES (COLUMNS 2 AND 3), AND NUMBER OF NAMED AIRWAY BRANCHES
NOT FOUND BY OTHER SEGMENTATION METHOD (COLUMNS 4 AND 5)

Fig. 9. Piecewise linear function for ! in cost function [see (6)]. Estimated
diameter is computed based on area from segmentation result (pixel-count in
re-sampled slice).

Alternatively, a validation scheme was developed that is

based on anatomical labeling and compares the segmentation result of the newly proposed algorithm with the result
of a region growing-based segmentation algorithm that was
previously published and that was specically developed for
airway tree segmentation [28]. Low-dose CT scans (120 kV,
50 mAs, voxel size 0.68 0.68 0.6 mm ) from 22 different
patients were used, and a human expert assigned as many
anatomical labels as possible to every segmentation result. A
total of 32 unique anatomical labels (Fig. 11) are commonly
used in human airway trees and the same level of labeling was
attempted for the reported validation. The 32 labels represent an
international standard used for anatomical labeling of airways
for bronchoscopic and other purposes [29]. For every CT scan
the two segmentation algorithms were compared based on the
number of discerned airway segments.
The newly proposed algorithm was run only once on every
CT dataset, using the same parameter settings for all datasets.
The region-grow algorithm was run 12 times on every dataset,
using different combinations of user-adjustable input parameters. The best region-grow result was hand-selected (by visually
inspecting all segmentation results and choosing the one result
with the highest number of branches and no signicant leaks).
This best region-grow result was then used for the anatomical
labeling.

Fig. 12 shows a typical comparison of two segmentation results.

B. Results
Table I lists the counts of named segments that were segmented by the two segmentation methods.
On average, the newly proposed segmentation method
anatomically named segments per tree
identied
, while the region-grow algonamed segments. The new method
rithm returned
statistically signicantly outperformed the region growing
. Across all tested trees there was a total
method
of 132 anatomically named segments that were segmented by
the new method, but missed by the region-grow algorithm. On
the other hand, there was only a total of 3 named segments that
have exclusively been segmented by the region-grow method.
In the most extreme single CT dataset, the new method found
23 named segments more than the region-grow method. In

1536

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO. 12, DECEMBER 2005

V. VALIDATION OF THE CROSS-SECTIONAL MEASUREMENTS

A. Method
The validation of the quantitative analysis method would ideally be based on in vivo scans. The problem with this approach is
that no accurate, reliable, and independent reference measurements are available (accuracy at voxel or subvoxel level, i.e.,
mm). Therefore, a physical phantom made of Plexiglas
tubes was used for the validation of the quantitative analysis.
The phantom contains 7 tubes of known diameters in the range
between 0.9819.25 mm. Since the lumen of the smallest tube
was invisible on the CT scans, the smallest tube was excluded
and the analysis was based on the remaining 6 tubes. The space
between the individual tubes was lled up with dried potatoakes, which closely resemble the texture of lung parenchyma
when scanned with a CT scanner. The phantom was scanned in
4 different angles of 0 , 5 , 30 , and 90 , rotated in the coronal
plane. Three different scan settings were used [low dose (120
kV, 50 mAs), regular dose (120 kV, 100 mAs), high dose (120
kV, 200 mAs)], all at a voxel size of 0.39 0.39 0.6 mm .
B. Results

Fig. 10. Segmentation result using the new method. Tree from the same CT
scan as used in Fig. 1.

Tables IIIV summarize the results from the measurements

on the Plexiglas phantom. The difference between nominal diameter
and the average measured diameter
is listed for
all combinations of tube size and scan angle . Mean and standard deviation values are given for all scan angles at a given
tube size (bottom two rows), as well as for all tube sizes at a
given scan angle (two rightmost columns).
It can be seen that the average absolute deviation from
the nominal diameter never exceeded 0.26 mm. The single
mm (Tube nr. 5
largest deviation was measured to be
of 9.50 mm, scanned at
with a nominal inner diameter
, regular dose). Comparing this with the voxel size
of 0.391 0.391 0.6 mm shows that the method works at
subvoxel accuracy. This is possible because the gray-image is
re-sampled before the border is detected (see Section III).
VI. DISCUSSION

Fig. 11. Airway tree with assigned labels. Labels refer to segments, but are
assigned to terminating branchpoint of respective segment. Drawing based on
[29].

contrast to that, the region-grow algorithm never identied

more than one extra named segment per tree in comparison
with the new method.
Fig. 14 shows the average count of anatomically named segments as a function of the generation number.

Compared with region growing-based airway tree segmentation, the newly proposed segmentation algorithm not only identies more airway segments (higher mean number of retrieved
segments), but it does so more consistently (smaller standard
deviation). This difference becomes especially apparent with
trees like the one depicted in Fig. 10, where the new method
returns a markedly better result. From Fig. 14 it can be seen
that the new segmentation method always returns all segments
of the rst 3 generations (zero standard deviation), whereas the
region-grow method misses branches of generation-numbers as
low as 2 (main bronchi) in some cases. It is also notable that the
new algorithm achieves this result without the need for the user
to hand-optimize parameters. In contrast, the parameters for the
region-grow algorithm were hand-optimized to get the best possible result.
The segmentation result is mostly unaffected by the choice
of the initial seedpoint. In some cases, minor leaks into the surrounding lung parenchyma may occur. This is the case if a leak

TSCHIRREN et al.: SEGMENTATION AND AIRWAY MORPHOLOGY ANALYSIS FROM LOW-DOSE CT SCANS

1537

Fig. 12. Segmentation resultlow-dose scan. (a) Region growing. Best possible result after hand-optimizing parameters. (b) Newly proposed algorithm. Fully
automated run without changing parameters.

MEASUREMENTS

ON

MEASUREMENTS

TABLE II
PLEXIGLAS PHANTOM. LOW DOSE SCAN (120
kV, 50 mAs)

TABLE III
PHANTOM. REGULAR DOSE
100 mAs)

ON PLEXIGLAS
SCAN (120 kV,

is solid, i.e., it does not have any holes and consequently it is

not recognized as a leak. However, this happens relatively rarely,
and if it does then such a leak is normally restricted to a rela-

TABLE IV
MEASUREMENTS ON PLEXIGLAS PHANTOM. HIGH DOSE SCAN (120 kV,
200 mAs)

tively small area. We never observed high-volume leaks such as

the ones often seen in region-grow-based segmentation results.
The segmentation method is sensitive to motion artifacts such
as heart beat induced bronchus movement, as well as airway
obstructions caused for example by mucus. If one of these artifacts causes an airway-lumen to be completely obstructed then
the segmentation of the affected airway branch will not be continued at this point (the same happens in region-grow-based algorithms). It is mostly small airways (at the sublobar level) that
are affected by this. We plan to address this problem in our future work.
The quantitative analysis achieves subvoxel accuracy on average. Only a very few isolated measurements show a deviation
of around 1 voxel from the nominal value.
It is interesting to note that no considerable difference
in the average accuracy can be observed between measure-

1538

IEEE TRANSACTIONS ON MEDICAL IMAGING, VOL. 24, NO. 12, DECEMBER 2005

Fig. 13. Diameters 6 standard deviation of named segments. Based on 22

normal subjects.

Fig. 14. Count of retrieved anatomically named branches, sorted by generation

number. Mean 6 standard deviation are reported. For the generations marked
with 3, the new method retrieves a statistically signicant higher count
of branches (p = 0:013; 0:037, and 0:004 for generations 4, 5, and 6,
respectively).

ments on airways that run perpendicular to the scan-plane

), and airways that run in-plane (phantom
(phantom at
). Other methods for the quantitative analysis of
at
airwaysnamely those that conduct measurements on the 2-D
scan-plane onlyshow difculties measuring in-plane or close
to in-plane airways.
A scan of a Plexiglas-and-potato-akes phantom is, of course,
not an ideal substitute for a scan of an in vivo human lung. It
would certainly be desirable to do a similar validation on in vivo
data. Unfortunately, no precise reference measurements (with
an accuracy of 0.5 mm or better) are available for in vivo data.
The phantom used here is the best currently available reference.

was tested on a total of 22 low-dose scans. The new segmentation algorithm proves to be considerably more robust than
region-grow-based airway segmentation algorithms, since parameters that have to be tuned by the user have been completely
eliminated. In many cases, the new algorithm outperforms
region growing-based segmentation methods. For many of the
scans region-grow-based methods provide usable results only
after several runs of the algorithm and considerable tuning of
parameters. The new airway-segmentation algorithm presented
here delivers good or very good results in all cases, after only
one single run of the algorithm per tree. In 2 out of the tested
22 cases, the new algorithm returned a somewhat inferior result
when compared to the region-grow-based algorithm. In these 2
trees, high generation airway segments were missing, although
the resulting trees were generally still of a good quality. A
total failure of the new algorithm, for example, with severe
leaks or missing lobes or sublobes, was not observed in any
of the test cases. Fig. 13 shows an application example of the
measurements algorithm, applied to the 22 low-dose scans.
The runtime of the newly proposed algorithm is kept down
to about 3 to 10 min (measured on a 1.2-GHz AMD Athlon
system) and depends on the size of the tree and the image
quality. Segmentation time was measured on a single-CPU
system. The segmentation algorithm is highly parallelizable.
Utilizing both CPUs of a symmetric multiprocessor (SMP)
system is expected to cut the segmentation time down to almost
half.
The proposed method for the quantitative analysis of airwaytree segments works fully in 3-D and performs the measurements in the original gray-scale volume for increased accuracy.
Information from anatomical labeling is used, which makes it
possible to perform measurements on specic anatomical segments named by the user. The algorithm was veried on a series
of high-resolution scans taken from a physical phantom. The
phantom contains Plexiglas tubes with known diameters ranging
from 1.98 mm to 19.25 mm. The validation showed that the proposed method delivers subvoxel accuracy for all scan-directions
(including in-plane airways).
The analysis of a complete airway tree scanned at the total
s (measured on a 1.2-GHz
lung capacity takes about 3 min
AMD Athlon system), depending on the size of the tree. This
includes the measurement of all anatomically named segments;
measurements are taken at every centerline point along the full
length of the segments.
ACKNOWLEDGMENT
The authors would like to thank M. Urschler, Technical University Graz, Austria, for developing the software tool that allowed the hand-labeling of the segmentation results.
REFERENCES

VII. CONCLUSION
A segmentation algorithm has been developed that works
fully in 3-D. It is able to detect leaks and prevent them as they
occur. The developed application is user-friendlythere are
no parameters that have to be tuned at runtime. The algorithm

[1] R. Chiplunkar, J. M. Reinhardt, and E. A. Hoffman, Segmentation and

quantitation of the primary human airway tree, presented at the SPIE
Conf.(Medical Imaging), San Diego, CA, 1997.
[2] T. Tozaki, Y. Kawata, N. Niki, H. Ohmatsu, R. Kakinuma, K. Eguchi, M.
Kaneko, and N. Moriyama, Pulmonary organs analysis for differential
diagnosis based on thoracic thin-section CT images, IEEE Trans. Nucl.
Sci., vol. 45, no. 12, pp. 30753082, Dec. 1998.

TSCHIRREN et al.: SEGMENTATION AND AIRWAY MORPHOLOGY ANALYSIS FROM LOW-DOSE CT SCANS

[3] K. Mori, Y. Suenaga, and J. Toriwaki, Automated anatomical labeling

of the bronchial branch and its application to the virtual bronchoscopy,
IEEE Trans. Med. Imag., vol. 19, no. 2, pp. 103114, Feb. 2000.
[4] T. Y. Law and P. A. Heng, Automated extraction of bronchus from 3D
CT images of lung based on genetic algorithm and 3D region growing,
Proc. SPIE (Medical Imaging), vol. 3979, pp. 906916, 2000.
[5] C. Pisupati, L. Wolf, W. Mitzner, and E. Zerhouni, Mathematical Morphology and Its Applications to Image and Signal Processing. Norwell, MA: Kluwer Academic, 1996, ch. Segmentation of
3D pulmonary trees using mathematical morphology, pp. 409416.
[6] F. Prteux, C. I. Fetita, P. Grenier, and A. Capderou, Modeling, segmentation, and caliber estimation of bronchi in high-resolution computerized tomography, J. Electron. Imag., vol. 8, no. 1, pp. 3645, 1999.
[7] C. I. Fetita and F. Prteux, Quantitative 3D CT bronchography, presented at the IEEE Int. Symp. Biomedical Imaging, Washington, DC, 1,
2002. (ISBI02).
[8] D. Bilgen, Segmentation and analysis of the human airway tree from
3D X-ray CT images, masters thesis, Dept. Biomed. Eng., Univ. Iowa,
Iowa City, 12, 2000.
[9] A. P. Kiraly, 3D image analysis and visualization of tubular structures,
Ph.D. dissertation, Dept. Comput. Sci. Eng., The Pennsylvania State
Univ. , University Park, 2003.
[10] D. Aykac, E. A. Hoffman, G. McLennan, and J. M. Reinhardt, Segmentation and analysis of the human airway tree from three-dimensional
X-ray CT images, IEEE Trans. Med. Imag., vol. 22, no. 8, pp. 940950,
Aug. 2003.
[11] M. Sonka, G. Sundaramoorthy, and E. A. Hoffman, Knowledge-based
segmentation of intrathoracic airways from multidimensional high resolution CT images, Proc. SPIE (Physiology and Function from Multidimensional Images) , vol. 2168, 1994.
[12] W. Park, E. A. Hoffman, and M. Sonka, Segmentation of intrathoracic
airway trees: A fuzzy logic approach, IEEE Trans. Med. Imag., vol. 17,
no. 8, pp. 489197, Aug. 1998.
[13] T. Kitasaka, K. Mori, H.-I. Hasegawa, Y. Suenaga, and J.-i. Toriwaki,
Extraction of bronchus regions from 3D chest X-ray CT images by
using structural features of bronchus, in Proc. Int Congr. Computer Assisted Radiology and Surgery (CARS) 2003, 2003, pp. 240245.
[14] T. Schlathlter, C. Lorenz, I. C. Carlsen, S. Renisch, and T. Deschamps,
Simultaneous segmentation and tree reconstruction of the airways for
virtual bronchoscopy, Proc. SPIE (Medical Imaging 2002: Image Processing), no. 2, pp. 103113, 2002.
[15] G. G. King, N. L. Mller, and P. D. Par, Evaluation of airways in
obstructive pulmonary disease using high-resolution computed tomography, Am. J. Respir. Crit. Care Med., vol. 159, no. 3, pp. 9921004,
1999.
[16] G. G. King, N. L. Mller, K. P. Whittall, Q.-S. Xiang, and P. D. Par,
An analysis algorithm for measuring airway lumen and wall areas from
high-resolution computed tomographic data, Am. J. Respiratory Crit.
Care Med., vol. 161, no. 2, pp. 574580, 2000.

1539

[17] S. A. Wood, E. A. Zerhouni, J. D. Hoford, E. A. Hoffman, and W.

Mitzner, Quantitative 3-D reconstruction of airway and pulmonary vascular trees using HRCT, in Proc. SPIE (Biomedical Image Processing
and Biomedical Visualization), vol. 1905, 1993, pp. 316323.
[18] J. M. Reinhardt, N. D. DSouza, and E. A. Hoffman, Accurate measurement of intrathoracic airways, IEEE Trans. Med. Imag., vol. 16, no. 12,
pp. 820827, Dec. 1997.
[19] J. M. Reinhardt, W. Park, E. A. Hoffman, and M. Sonka, Intrathoracic
airway wall detection using graph search with CT scanner PSF information, Proc. SPIE (Medical Imaging), vol. 3033, pp. 93101, Feb. 2328,
1997.
[20] F. Prteux, C. I. Fetita, and P. Grenier, Modeling, segmentation, and caliber estimation of bronchi in high-resolution computerized tomography,
Proc. SPIE (Statistical and Stochastic Methods in Image Processing II),
vol. 3167, no. 7, pp. 5869, 1997.
[21] O. I. Saba, E. A. Hoffman, and J. M. Reinhardt, Computed tomographic-based estimation of airway size with correction for scanned
plane tilt angle, Proc. SPIE (Physiology and Function from Multidimensional Images), vol. 3978, no. 2, 2000.
[22] R. Wiemker, T. Blaffert, T. Blow, S. Renisch, and C. Lorenz, Automated assessment of bronchial lumen, wall thickness, and bronchoarterial diameter ratio of the tracheobronchial tree using high-resolution
CT, in Proc. CARS, 2004, pp. 967972.
[23] J. K. Udupa and S. Samarasekera, Fuzzy connectedness and object definition: Theory, algorithms, and applications in image segmentation, in
Graphics Models Image Process , vol. 58, 1996, pp. 246261.
[24] G. T. Herman and B. M. Carvalho, Multiseeded segmentation using
fuzzy connectedness, IEEE Trans. Pattern Anal. Mach. Intell., vol. 23,
no. 5, pp. 460474, May 2001.
[25] K. Palgyi, J. Tschirren, and M. Sonka, Quantitative analysis of threedimensional tubular tree structures, Proc. SPIE (Medical Imaging), vol.
5032, no. 2, p. 277, Feb. 2003.
[26] M. Sonka, V. Hlavac, and R. Boyle, Image Processing, Analysis, and
Machine Vision. Pacic Grove, CA: Brooks/Cole, 1998.
[27] M. Sonka, G. K. Reddy, M. D. Winniford, and S. M. Collins, Adaptive approach to accurate analysis of small-diameter vessels in cineangiograms, IEEE Trans. Med. Imag., vol. 16, no. 2, pp. 8795, Feb. 1997.
[28] A. P. Kiraly, W. E. Higgins, G. McLennan, E. A. Hoffman, and J. M.
Reinhardt, Three-dimensional human airway segmentation methods
for clinical virtual bronchoscopy, Acad. Radiol., vol. 9, no. 10, pp.
11531168, 2002.
[29] E. A. Boyden, Segmental Anatomy of the Lungs. New York: McGrawHill, 1955.